factor-mining

Name: factor-mining
Author: minihellboy

95stars28forks

How to install factor-mining?

minihellboy/factorminer/factor-mining

$npx -y skills add minihellboy/factorminer --skill factor-mining

Installs into the current project.

Summary

Discover alpha factors by running the FactorMiner research engine — the paper-faithful Ralph loop or the enhanced Helix loop (causal validation, regime conditioning, multi-specialist debate, canonicalization). Use to generate a new factor library from a validated dataset. Trigger

Files · 1

View on GitHub

SKILL.md

1	# Factor Mining
2
3	This skill runs FactorMiner's self-evolving discovery loop: it retrieves memory priors, proposes candidate factor formulas with an LLM, evaluates them, and admits the survivors to a factor library.
4
5	See `references/loop-architecture.md` for the stage-by-stage loop design and `references/dsl-operators.md` for the factor-formula operator vocabulary.
6
7	## Choosing the loop
8
9	\| Use \| When \|
10	\|---\|---\|
11	\| `mine` (Ralph loop) \| Default. Paper-faithful Algorithm 1 — retrieve, generate, evaluate, admit, evolve memory. \|
12	\| `helix` (Helix loop) \| When you want Phase 2 features: do-calculus causal validation, regime-conditional evaluation, multi-specialist debate generation, or SymPy canonicalization. Drop-in superset of Ralph. \|
13
14	## Workflow
15
16	### 1. Confirm prerequisites
17
18	The dataset must already pass `factor-data` validation. Confirm the iteration budget — mining cost scales with `iterations × batch-size`.
19
20	### 2. Run the Ralph loop
21
22	```bash
23	factorminer -o output/run1 mine \
24	--data path/to/market_data.csv \
25	--iterations 40 --batch-size 16 --target 30
26	```
27
28	- `--iterations` — maximum mining iterations (the loop also stops early once `--target` factors are admitted).
29	- `--batch-size` — candidate factors proposed per iteration.
30	- `--target` — desired library size.
31	- `--resume path/to/factor_library.json` — continue a previous run.
32	- `--mock` — synthetic data + mock LLM, no API calls. Use only for smoke tests.
33
34	### 3. Or run the Helix loop
35
36	```bash
37	factorminer -o output/run1 helix \
38	--data path/to/market_data.csv \
39	--iterations 40 --batch-size 16 --target 30 \
40	--causal --regime --debate --canonicalize
41	```
42
43	Each `--feature / --no-feature` flag overrides the config; omit a flag to keep the config default. Phase 2 features cost extra compute and LLM calls — enable the ones the research question needs.
44
45	### 4. Inspect the result
46
47	```bash
48	factorminer session inspect output/run1 --json
49	```
50
51	Report library size, iteration count, and yield rate. The factor library is written to `output/run1/factor_library.json`; the run log to `session_log.json`.
52
53	## Guardrails
54
55	- Mining proposes formulas; it does not prove them. Always follow with `factor-evaluation` on the held-out split.
56	- A low yield rate usually means thresholds are too strict for the dataset, not that the data is bad — tune `ic_threshold` / `correlation_threshold` in config, do not silently relax them in a report.
57	- `--mock` output is never a research result; never present mock metrics as real.
58
59	## MCP alternative
60
61	When the FactorMiner MCP server is connected, `mine_factors` and `helix_mine` expose the same workflow as tools, returning a structured session summary directly.

People who liked this also viewed

6 picks

Type

factor-mining

95stars28forks

View on GitHub

How to install factor-mining?

minihellboy/factorminer/factor-mining

$npx -y skills add minihellboy/factorminer --skill factor-mining

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# Factor Mining
2
3	This skill runs FactorMiner's self-evolving discovery loop: it retrieves memory priors, proposes candidate factor formulas with an LLM, evaluates them, and admits the survivors to a factor library.
4
5	See `references/loop-architecture.md` for the stage-by-stage loop design and `references/dsl-operators.md` for the factor-formula operator vocabulary.
6
7	## Choosing the loop
8
9	\| Use \| When \|
10	\|---\|---\|
11	\| `mine` (Ralph loop) \| Default. Paper-faithful Algorithm 1 — retrieve, generate, evaluate, admit, evolve memory. \|
12	\| `helix` (Helix loop) \| When you want Phase 2 features: do-calculus causal validation, regime-conditional evaluation, multi-specialist debate generation, or SymPy canonicalization. Drop-in superset of Ralph. \|
13
14	## Workflow
15
16	### 1. Confirm prerequisites
17
18	The dataset must already pass `factor-data` validation. Confirm the iteration budget — mining cost scales with `iterations × batch-size`.
19
20	### 2. Run the Ralph loop
21
22	```bash
23	factorminer -o output/run1 mine \
24	--data path/to/market_data.csv \
25	--iterations 40 --batch-size 16 --target 30
26	```
27
28	- `--iterations` — maximum mining iterations (the loop also stops early once `--target` factors are admitted).
29	- `--batch-size` — candidate factors proposed per iteration.
30	- `--target` — desired library size.
31	- `--resume path/to/factor_library.json` — continue a previous run.
32	- `--mock` — synthetic data + mock LLM, no API calls. Use only for smoke tests.
33
34	### 3. Or run the Helix loop
35
36	```bash
37	factorminer -o output/run1 helix \
38	--data path/to/market_data.csv \
39	--iterations 40 --batch-size 16 --target 30 \
40	--causal --regime --debate --canonicalize
41	```
42
43	Each `--feature / --no-feature` flag overrides the config; omit a flag to keep the config default. Phase 2 features cost extra compute and LLM calls — enable the ones the research question needs.
44
45	### 4. Inspect the result
46
47	```bash
48	factorminer session inspect output/run1 --json
49	```
50
51	Report library size, iteration count, and yield rate. The factor library is written to `output/run1/factor_library.json`; the run log to `session_log.json`.
52
53	## Guardrails
54
55	- Mining proposes formulas; it does not prove them. Always follow with `factor-evaluation` on the held-out split.
56	- A low yield rate usually means thresholds are too strict for the dataset, not that the data is bad — tune `ic_threshold` / `correlation_threshold` in config, do not silently relax them in a report.
57	- `--mock` output is never a research result; never present mock metrics as real.
58
59	## MCP alternative
60
61	When the FactorMiner MCP server is connected, `mine_factors` and `helix_mine` expose the same workflow as tools, returning a structured session summary directly.

People who liked this also viewed

6 picks

Type